Who responds more to environmental amenities and dis-amenities?

Land Use Policy 62 (2017) 151–158

Contents lists available at ScienceDirect

Land Use Policy journal homepage: www.elsevier.com/locate/landusepol

Who responds more to environmental amenities and dis-amenities? Darshana Rajapaksa a , Clevo Wilson b , Viet-Ngu Hoang b , Boon Lee b , Shunsuke Managi a,b,∗ a b

Urban Institute, Department of Urban and Environmental Engineering, Faculty of Engineering, Kyushu University, 744, Motooka Nishi-ku, Fukuoka, Japan QUT Business School, Queensland University of Technology, Level 8, Z Block, Gardens Point, 2 George St, Brisbane QLD 4000, Australia

a r t i c l e

i n f o

Article history: Received 5 September 2016 Received in revised form 14 November 2016 Accepted 30 December 2016 JEL classification: Q51 Q54

a b s t r a c t It is hypothesised that different property sub-markets react to flood risk information, floods and environmental factors differently. To test this hypothesis this research uses spatial quantile regression and quasi-experimental techniques to examine property sub-market behaviour in response to availability of flood risk information and actual flood. This new contribution to the literature is based on the use of the mapping of flood risk areas in 2009 and the 2011 flooding of Brisbane, Australia, as a case study. The results show that the impact of flood risk and actual flood on property markets varies between different sub-markets. They therefore confirm the existence of property sub-markets based on property and environmental characteristics and suggest the need for differentiate mitigation policies. © 2017 Elsevier Ltd. All rights reserved.

Keywords: Flood Hedonic property price Sub-markets Spatial econometrics

1. Introduction Whether all property buyers/sellers (representing different submarkets) consider environmental amenities and dis-amenities in a similar manner is questionable. To answer this question, this research investigates the variation in impact of the release of flood risk information and actual flood incidence on different property sub-markets. We hypothesise that the property market can be categorised by market value of the property and which assumes rich people tend to buy high-value properties whereas poor people tend to buy low-value properties. The hypothesis is tested, first using quantile regression of hedonic price (HP) analysis and then by a two-stage quantile regression (2SQR) analysis to correct spatial autocorrelation. Finally, the quasi-experimental technique is combined to distinguish the impact of the release of flood risk information and actual flood incidence across different submarkets. Property characteristics, namely quality and size of the property are key determinants of property prices. But just as important are the environmental factors such as neighbourhood amenities, open spaces and greenery areas which significantly contribute to

∗ Corresponding author at: Urban Institute, Department of Urban and Environmental Engineering, Faculty of Engineering, Kyushu University, 744, Motooka Nishi-ku, Fukuoka, Japan. E-mail address: [email protected] (S. Managi). http://dx.doi.org/10.1016/j.landusepol.2016.12.029 0264-8377/© 2017 Elsevier Ltd. All rights reserved.

market clearance prices (Irwin, 2002; Kong et al., 2007; Hibiki and Managi, 2011). Landry and Hindsley (2011) have observed that properties close to a beach have higher value than similar properties farther away. Furthermore, environmental dis-amenities – floods, bushfire, and environmental pollution – exert a negative impact on property values. For example, Gawande et al. (2013) identified the risks from nuclear waste transportation in Mexico as having a negative impact on property prices. The evidence from many HP studies show the impact of such factors on property prices vary temporally and spatially (Lamond et al., 2010). This is largely due to the inherent nature of geographical distribution and heterogeneity in property markets which naturally create sub-markets. For example, some properties are close to natural geographical formations (such as waterfronts, natural forests, views) or manmade infrastructure (such as playgrounds, parks, recreational places, religious establishments, the central business district (CBD), shopping centres, schools, roads and transport). Hence by taking into account property sub-markets, results from HP analysis becomes more precise and reliable (Bourassa et al., 1999). It has been generally accepted that geographical boundaries or administrative boundaries are the best approximations for submarkets because, within a given geographical area, the property market is more homogeneous. However according to Mak et al. (2010) property sub-markets can be observed even within a single condominium. Other than structural, geographical and environ-

152

D. Rajapaksa et al. / Land Use Policy 62 (2017) 151–158

mental factors, part of this heterogeneity in the product (property) can be attributed to behavioural and social factors. According to Hoshino (2011), residential preferences are also heterogeneous. For instance, well off people tend to buy properties in an area where resource and environmental factors have value. Given these facts and complexities, behavioural factors can be a more appropriate means to distinguish sub-markets. For example, property sub-markets can be closely related to the income group of property residents given that high-value properties are largely owned by high-income groups whereas low-value properties are largely owned by low-income groups. Similarly, these groups value environmental amenities and dis-amenities differently. Hence, in the current study the definition of a property sub-market is based on property values and which, it is assumed, also relate to household income and other associated characteristics. Although a great deal of theoretical and empirical work has been carried out on various aspects of flood impacts on the property market, much less attention has been paid to the behaviour of different property sub-market affected by natural disasters. Therefore, the focus of this paper is to examine, firstly, whether there are similarly marked differences in property sub-markets subject to natural hazards and floods and, secondly, to the release of flood risk information. In particular this research seeks to answer the following research questions which are so far unaddressed in the literature: (1) are low-value properties affected more than high-value properties? (2) which sub-markets (low-value properties or high- value properties) are more price responsive to the availability of flood risk information? (3) which creates more a more negative price impact: availability of risk information or the experience of an actual flood event? In this study spatial quantile regression and quasi-experimental analysis are applied to two situations – making available flood risk information and an actual flood – as a natural experiment. The study finds that both the availability of risk information and an actual flood impacts on property markets differently between low-value properties and high-value property. This research, therefore, makes an important contribution to the existing literature on the effects on property markets of natural disasters, by exploring the impacts of public risk information provision and actual natural disasters across different property sub-market. The rest of this article is organised as follows. Section 2 briefly reviews the different empirical approaches used to categorise property sub-markets and analyse property market behaviour. The methodology employed in this paper is discussed in Section 3. Empirical results are presented in Section 4 and Section 5 concludes with some brief recommendations. 2. Literature review HP analysis is widely used in valuing environmental amenities and dis-amenities. For example, Kong et al. (2007) and Nicholls and Crompton (2005) have shown the manner in which physical access or views of green space has a positive and significant impact on determining house prices. Similarly, Gopalakrishnan et al. (2011) observed that being close to a water view had a similar impact on house prices. Among the extensive literature on environmental dis-amenities, a number of studies have investigated the behavior of the property market in relation to natural disasters1 including flooding and the presence of flood plains

1 Examples of other dis-amenities include impact of air pollution (see, for example, Zabel and Kiel, 2000; Kim et al., 2003), telecommunication tower (see, for example, Filippova and Rehm, 2011), nuclear waste (see, for example, Gawande et al., 2013), noise (see, for example, Duarte and Tamez, 2009), other characters (Thompson et al., 2012).

(Rajapaksa et al., 2016; Bin and Landry, 2013; Petrolia et al., 2013; Rambaldi et al., 2013; Samarasinghe and Sharp, 2010; Lamond et al., 2010; Zhai and Fukuzono, 2003; Fridgen and Shultz, 1999). Some studies have shown that flood risk discounts property values (see, for example, Rambaldi et al., 2013; Lamond et al., 2010) whereas, others have indicated a negative impact following an actual flood incidence (see, for example, Bin and Landry, 2013). In a study of New Zealand property sales data, Samarasinghe and Sharp (2010) showed a significant negative impact on property prices from being located in flood prone areas. A study of the recent effects of hurricanes in North Carolina by Bin and Landry (2013) employing a difference-in-differences (DID) framework found the presence of a risk premium ranging between 6% and 20% for properties in the flood prone zone. However, a related study in the UK found that the impact of floods varies temporally and spatially (Lamond et al., 2010). A qualitative analysis of price behaviour of the Brisbane property market showed that floods create negative effects on the average listing price (Eves and Wilkinson, 2014). However, other than a recent study by Zhang (2016), none of these studies have investigated the variation in negative impact across different sub-markets. Using quantile regression analysis Zhang (2016) found that the flood risk impacts on housing market differently. However, this study is different as we combined spatial quantile regression and quasi-experimental analysis to compare two events across different sub-markets. Reviewing 125 research articles, Sirmans et al. (2005) show the inconsistency of parameter estimation, even among commonly used property related variables in HP analysis. This is possible due to the highly heterogeneous nature of the property market and the existence of property sub-markets. Buyers do not tend to bid for properties in a number of differentiated sub-markets but rather for similar sub-markets given that different property sub-markets behave differently. Thus the property demand and supply structure differs across different market segment (Freeman, 1993). According to Farmer and Lipscomb (2010), households compete with each other within their own sub-market. Furthermore, within a submarket, properties are more homogeneous and hence estimations are more precise (Bourassa et al., 1999). As suggested by Michaels and Smith (1990), separate HP functions for different market segments will produce more precise estimations of the relationship between property characteristics and property price than a single HP function. In addition, Miron (1995) showed that hedonic rental prices vary from city to city in Canada. Property sub-markets and their characteristics have been extensively researched in real estate and marketing literature. Bourassa et al. (1999) in adopting the K-means clustering procedure and principle component analysis to identify property sub-markets in Sydney and Melbourne found that the HP estimation for sub-markets is more appropriate than for the whole market. As Wilhelmsson (2004) suggests, the HP predictive power can be improved while reducing spatial dependency by identifying property sub-markets. A number approaches have been used to distinguish different housing sub-markets. Most have used geographical boundaries, administrative boundaries, school boundaries and census boundaries. Dale-Johnson (1982) used factor analysis whereas Wilhelmsson (2004) employed a cluster analysis in clustering Swedish property markets. In contrast to statistical approaches, Dorsey et al. (2010) used zip codes to define property sub-markets. As noted, if sub-markets are ignored in the HP analysis the estimation coefficients become biased. Another approach to sub-market classification is to use income information. Gayer (2000) provides empirical evidences for the existence of behavioural differentiation of property sub-markets clustered by income neighbourhoods. The research observed that high income neighbourhoods placed a higher valuation on environmental risk reduction than low income neighbourhoods.

D. Rajapaksa et al. / Land Use Policy 62 (2017) 151–158

153

Fig. 1. Flood map of Brisbane.

McCluskey and Rausser (2003) also showed that property markets can be segmented into upper-priced range properties and lowerpriced range properties or in terms of high-income and low-income groups. In particular, housing characteristics were valued differently between the two segments. Similarly, it may be assumed that environmental factors such as flood-risk are valued differently across the sub-markets. We aim to provide one of the first empirical examinations to test these assumptions using the 2011 Brisbane flood data and release of Australian flood risk information in 2009. Providing evidence of sub-markets in the analysis of flood impacts on property markets is important given reliable empirical results could provide more meaningful input into the design and implementation of differentiated policies assisting flood victims (Prasad and Richards, 2008). For example, the improved estimation of flood damage and appropriate policies can be achieved through identification of different sub-markets. According to Merz et al. (2004), when estimating direct flood damage, the variability in estimation can be minimised by more appropriate classifications of affected properties − which is another approximation for identifying sub-markets.

3. Econometric methods 3.1. Quantile regression analysis Given the price of a property is used to distinguish a sub-market and also between the dependent variables in HP models, quantile regression analysis is preferred to ordinary least square (OLS) estimates. More specifically this is due to two reasons. First, it allows for different a relationship between property prices and their covariates at a different quantile of the asymmetric distribution of property prices. Second, the quantile regression analysis

reduces the effect of outliers (McMillen and Thorsnes, 2006) and unobserved heterogeneity (Koenker, 2005). Since the seminal work of Koenker and Bassett (1978), quantile regressions have been widely used in HP estimation (see, for example, Ebru and Eban, 2011; Farmer and Lipscomb, 2010; Heintzelman, 2010; Kostov, 2009; Zietz et al., 2008 and McMillen and Thorsnes, 2006). Zietz et al. (2008) uses quantile regression to show that the buyers of higher-priced houses value square footage and number of bedrooms differently from the buyers of lowerpriced houses. Kostov (2009) applies quantile regression to the hedonic land price model in Northern Ireland to study the differences of segmented land markets. Kuethe and Keeney (2012) use quantile regression techniques to analyse the impact of animal agriculture on property prices. These studies support the view that quantile regression analysis is more appropriate for segmented property markets. The linear regression model of the HP function can be expressed as: pi = ∝i +



ˇi xi + εi

(1)

where, pi is the property price, xi the independent variables, ∝i the constant term and εi the error term. Eq. (1) describes the conditional mean of property prices. The conditional mean of Pi given xi can be expressed as: E (pi |xi ) = ˛i + ˇi xi Following the literature (Hao and Naiman, 2007 and Koenker and Bassett, 1978), quantile regression model can be expressed as: (q)

pi = ˛i

+



(q)

(q)

ˇi xi + εi

(2)

154

D. Rajapaksa et al. / Land Use Policy 62 (2017) 151–158

Table 1 Selected variables. Variable

Description

LR PRICE BEDRM BATHRM GARAGE FLOORAREA WALL BRICK WALL WOOD

House sale price adjusted to 2013 quarter 1 price Number of bedrooms Number of bathrooms Number of garage spaces Total floor are in square meters Dummy variable for wall construction material: 1 if brick, 0 otherwise Dummy variable for wall construction material: 1 if wood, 0 otherwise (The omitted category is mixed wood and brick) Dummy variable for floor construction material: 1 if carpet, 0 otherwise Dummy variable for floor construction material: 1 if tile, 0 otherwise Dummy variable for floor construction material: 1 if wood, 0 otherwise (The omitted category is mixed) Dummy variable for roof construction material: 1 if tile, 0 otherwise Dummy variable for roof construction material: 1 if asbestos, 0 otherwise (The omitted category is galvanised iron) Proxy variable for type of house: Number of stories in the house Dummy variable: 1 if there is a swimming pool, 0 otherwise Dummy variable: 1 if a garden is present and 0 otherwise. (The presence of well-maintained gardens at the time of selling is considered as a presence of a garden) Dummy variable: 1 if property is on the water front, 0 otherwise Dummy variable for the flood affected property: 1 if the property is flood affected, 0 otherwise Distance in meters to the Brisbane river or water stream Medium household income in the mesh block (smallest census block) Direct distance in meters to the closest park Direct distance in meters to the railway track Direct distance in meters to the highway Distance in meters to the closest shopping centre Distance in meters to the closest primary school Distance in meters to the central business district Dummy variable for release of flood risk information in 2009: 1 if sold after, 0 otherwise Dummy variable for flood 2011: 1 if sold after, 0 otherwise

FLOOR CARP FLOOR TILE FLOOR WOOD ROOF TILE ROOF ASBES

HOUSETYPE STORIES POOL GARDEN WATERFRONT FLOOD DIS RIVER INCOME DIS PARK DIS RAIL DIS HIGWAY DIS SHOP DIS PSCHOL DIS CBD BFAF INFOR BFAF FLOOD

where, 0 < q < 1 indicates the proportion of the population below the qth quantile as shown below: (q)

Qq (pi |xi ) = ␣i

+



(q)

␤i x i

(3)

The quantile regression is based on the minimisation of a weighted sum of the absolute deviations; min

 k bj

j=0

 i

 k

|pi −

bj xj,i |hi

extended to include a spatial lag effect (Kostov 2009; Su and Yang, 2007). In this study, the 2SQR method is used. First, an HP analysis is carried out. The linear spatial lag model is defined as:

(4)

j=0

where, pi is the dependant variable at observation i, xj,i, and the ith regressor variable at observation i, and bj is a vector of the coefficient estimate of the model’s jth regression coefficient. The weighted hi is defined as:hi = 2q if the residual for the ith observation is strictly positive or hi = 2 − 2q if the residual for the ith observation is negative or zero.

p = Wp + ␣ +



␤i xi + 

(5)

where, p is the property price, ␳ the spatial lag parameter, Wp the spatial lagged dependent variable, x the matrix of independent variables, ˇ the parameters of x and ␧ the vector of error terms. The spatial dependency is examined using Moran’s I and LM statistics3 and, as noted, a spatial lag effect is incorporated in the quantile regression analysis. In contrast to the spatial HP model, a spatial quantile can be specified as: p =  () Wp + ␣ +



␤i () xi + 

(6)

where,  is the corresponding quantile of the p, () the property price in the  th quantile and ˇ() quantile specific parameters.

3.2. Spatial quantile regressions

3.3. Quasi-experimental analysis

The estimates of the parameters in the quantile regression of HP analysis may not be efficient due to the spatial dependence of property price. Current literature has shown that ignoring spatial dependence can yield biased estimation (see, for example, Anselin, 1988). At least two methods have been proposed to incorporate spatial effects in quantile regression analysis. Kim and Muller (2004) proposed a two stage quantile regression (2SQR) analysis while Chernozhukov and Hansen (2008) proposed an instrumental variable quantile regression (IVQR).2 IVQR was then further

Natural disasters can be considered as a natural experiment and used to compare different disaster events over time and in which DID models have been employed in the literature (see, Rajapaksa et al., 2016; Bin and Landry, 2013 and Gawande et al., 2013). For

2 Kostov (2009) has reviewed the advantages and disadvantages of both approaches and concluded that both approaches create concerns about the endogeneity issue. The spatially lagged independent variables are considered as

instruments and the standard error is estimated using the bootstrapping procedure. The 2SQR is easier to estimate whereas in estimating IVQR there is a need to define several steps. 3 Moran’s I and LM statistics indicate the presence of a spatial lag effect. Moran’s I can be estimated using the following equation: Morans I =

¯ j − x) ¯ i j wij (xi − x)(x n 2 i j wij ¯ j (xi − x)

D. Rajapaksa et al. / Land Use Policy 62 (2017) 151–158

155

Table 2 Descriptive statistics.

R PRICE BEDRM BATHRM GARAGE FLOORAREA HOUSETYPE FLOOR CARP FLOOR TILE FLOOR WOOD ROOF TILE ROOF ASBES WALL BRICK WALL WOOD POOL GARDEN STORIES WATERFRONT INCOME DIS PARK DIS HIGWAY DIS SHOP DIS RAIL DIS CBD DIS RIVER FLOOD

Mean

Min

Max

Mean values of different quantile

13.2425 3.4571 1.7245 1.5885 649.29 5.7058 0.1872 0.0652 0.3245 0.4299 0.2883 0.4486 0.3075 0.1710 0.7659 1.2499 0.0320 1687.06 597.39 1976.74 1748.89 1407.82 10084.56 2151.05 0.2158

11.66 1 1 0 119 1 0 0 0 0 0 0 0 0 0 0 0 614 10 50 500 10 6300 100 0

15.75 7 5 6 5640 8 1 1 1 1 1 1 1 1 1 3 1 3291.13 1900 4100 3220 4620 15000 6700 1

Q1 12.7234 3.2785 1.4195 1.3977 608.90 5.5822 0.1862 0.0621 0.1879 0.3926 0.1460 0.3809 0.1896 0.0671 0.7550 1.1074 0.0034 1313.28 698.69 2077.85 1732.75 2683.14 12528.19 3854.53 0.1091

instance, the impact of flood can be considered as the difference between the property prices of affected and non-affected properties (that is, treatment effect) and which fundamentally cannot be observed (Khandker et al., 2010). In experimental analysis, two similar properties, one with treatment, are compared before and after the treatment. Here, two variables capturing time and events (treatment effects)4 are included into Eq. (6): p =  () Wp + ␣ +



␤i () xi + ∝t () dt df + 

(7)

where, dt represent the time period and df the treatment. 4. Empirical study 4.1. Study area and data description Brisbane is the capital of the state of Queensland in Australia. This city has experienced 6 major floods in the last two centuries − in 1841, 1890, 1893, 1931, 1974 and 2011.5 Of these, the Brisbane floods of 2011 was recorded as one of the worst natural disasters in Australia’s recent past. Brisbane, therefore, provides an ideal site for this type of empirical investigation not only because of its history of flooding and its major impact on property markets, but also because of the availability of flood related data. During the January 2011 flood, the Brisbane water level peaked at 4.46 m and caused damage greater than the previous record flood in 1974 (Eves and Wilkinson, 2014). Before the 2011 flood, the Brisbane City Council (BCC) mapped flood vulnerable areas and made it available to the public in 2009. This paper considers both events – the release of flood risk information (October 2009) and the actual

4 The release of flood-risk map in 2009 and actual floods in 2011 were considered as time demarcations (before and after release of flood-risk information and before and after actual flood) and flood affected and non-affected properties consider as treatment. In this research two time dummies (dt : BFAF INFOR and BFAF FLOOD) and one treatment (df : FLOOD) are employed. In quasi experimental analysis (DID estimation) D1 show the impact of release of flood information and D2 show the impact of actual flood. 5 Brisbane River Flood History (http://www.brisbane-australia.com/brisbaneriver-flood-history.html).

Q2 13.0421 3.2609 1.4883 1.5452 613.39 5.4666 0.2341 0.0702 0.3344 0.4465 0.2174 0.3495 0.3227 0.0853 0.7358 1.1421 0.0033 1436.58 664.85 1772.58 1922.88 1634.95 11285.28 2705.85 0.1706

Q3 13.3044 3.4361 1.6900 1.5937 615.98 5.6953 0.2067 0.0701 0.3958 0.5009 0.3468 0.5149 0.3520 0.1121 0.7443 1.2557 0.0123 1743.32 547.27 2001.31 1765.04 741.59 9148.86 1363.92 0.2715

Q4 13.9217 3.8638 2.3155 1.8241 760.58 6.0897 0.1207 0.0586 0.3845 0.3810 0.4500 0.5552 0.3690 0.4241 0.8293 1.5017 0.1103 2274.02 473.07 2059.14 1570.17 519.02 7256.72 603.45 0.3172

flood (January 2011) and distinguishes the effects of both events across different-valued properties. As shown in Fig. 1, nearly 35 suburbs within the BCC were partly or completely affected along the Brisbane river bank. These suburbs have different demographics in terms of income, socio-economic characteristics and suburb neighbourhood characteristics. Some suburbs are close to the central business district (CBD) while others are some distance away. Some suburbs are substantially richer in natural and manmade characteristics (such as greenery areas and parks) than others. Overall then, suburb characteristics are highly heterogeneous. For this study, the first stage of the data collection involved the selection of appropriate suburbs after a detailed examination of the post-flood maps. Four flood-affected suburbs were selected. Two suburbs, Oxley and Durack, are a considerable distance from the CBD and the Brisbane river but in a low elevation region whereas Chelmer and Graceville are close to the river and have comparatively higher median income households. We considered only single dwelling property transaction data for the period 2006–2013. A final sample of 2345 observations was used in the econometric analysis. Property transaction data were drawn from the RP DATA information services (www.rpdata.com. au) while GIS techniques were used to construct other variables from various databases published by public organisations (road networks, stream networks, and land use maps). Table 1 describes the variables selected. The dependent variable, property price (log value – adjusted for inflation) was regressed with structural variables (e.g. number bedrooms, number of bathrooms), environmental variables (e.g. waterfront properties), socio-economics characteristics (e.g. median income, access to transport) and flood related variables (distance to the river, flood affected). Descriptive statistics of four quantiles for the selected variables are presented in Table 2. The mean value for number of bedroom for the entire sample is 3.8 and 1.7 for bathrooms. Distance to river varies from 100 m to 6700 m with a mean value of 2,152 m. Property characteristics also vary across the four quantiles: the mean value for number of bathrooms is higher for higher priced quantiles (2.3) compared to lower priced quantiles (1.4). The average medium income (weekly) shows a clearly increasing trend when related to

156

D. Rajapaksa et al. / Land Use Policy 62 (2017) 151–158

Table 3 HP analysis without considering spatial autocorrelation. OLS CONST BEDRM BATHRM GARAGE FLOORAREA HOUSETYPE FLOOR CARP FLOOR TILE FLOOR WOOD ROOF TILE ROOF ASBES WALL BRICK WALL WOOD POOL GARDEN STORIES WATERFRONT INCOME DIS PARK DIS HIGWAY DIS SHOP DIS RAIL DIS CBD DIS RIVER FLOOD BFAF INFOR BFAF FLOOD D1 D2 N Adj R2 Pseudo R2

13.1872**** 0.04297**** 0.10621**** 0.03642**** 0.00034**** 0.00846* 0.00803 0.00448 0.01763* −0.01780 0.00672 −0.02976** 0.00360 0.13213**** 0.06449**** 0.07474**** 0.47504**** 0.00016**** −0.00005*** −0.00002* 0.00001 −0.00003** −0.00011**** 0.00007**** −0.07147**** 0.07045**** −0.07292**** −0.01181 −0.13601**** 2345 0.73

Quantile regression analysis 0.1 13.0476**** −0.00120 0.09202**** 0.03347*** 0.00026**** −0.01827 −0.00842 −0.12338*** 0.03547 −0.00902 −0.01991 −0.03574 0.00301 0.12051**** 0.10242**** 0.059605** 0.22821*** 0.00017**** 0.00001 −0.00003 −0.00004 0.00002 −0.00007**** −0.00001 −0.07566* 0.11199*** −0.01169 0.07314 −0.13825*** 2345

0.25 12.9552**** 0.02630*** 0.09294**** 0.04030**** 0.00025**** 0.00991 −0.00024 −0.00763 0.01086 −0.02059 −0.00988 −0.01066 0.02948 0.10868**** 0.08704**** 0.08281**** 0.3236**** 0.00017**** 0.00001 −0.000025* −0.00002 −0.00004* −0.00008**** 0.00004* −0.04821** 0.07084**** −0.07619**** 0.01231 −0.11018*** 2345

0.5 13.1698**** 0.05694**** 0.09810**** 0.0246**** 0.00036**** 0.01855*** −0.00503 0.01230 0.02038* −0.02298* 0.00816 −0.03527** −0.01306 0.10027**** 0.05246**** 0.07774**** 0.45093**** 0.00015**** −0.00008**** −0.00001 0.00000 −0.00008**** −0.00010**** 0.000098**** −0.07253**** 0.03516*** −0.09864**** −0.00752 −0.10703**** 2345

0.75 13.26865**** 0.05842**** 0.12307**** 0.03308**** 0.00049**** 0.01839*** −0.00671 −0.01926 0.01426 −0.03844*** 0.00490 −0.04700*** −0.03534** 0.13961**** 0.03929*** 0.09259**** 0.60899**** 0.00012**** −0.00009**** 0.00001 0.00004*** −0.00010**** −0.00012**** 0.00012**** −0.07247**** 0.04877*** −0.10783**** −0.09339*** −0.14762**** 2345

0.9 13.0352**** 0.06399**** 0.13569**** 0.02846*** 0.00053**** 0.02519*** −0.01948 −0.00532 −0.01507 −0.04907*** 0.00163 −0.01466 0.01874 0.17093**** 0.04029*** 0.09928**** 0.59036**** 0.00013**** −0.00007*** 0.00003** 0.00008**** −0.00006*** −0.0001**** 0.00007*** −0.0781*** 0.08104**** −0.09661**** −0.11244*** −0.17563**** 2345

0.33

0.41

0.51

0.59

0.64

Note: ****, ***, ** and * denote coefficients that are significant at 1,5,10 and 20% levels of significance, respectively.

the price of a house (AUS $2274 for higher quantiles and AUS$1313 for lower quantiles). The high correlation (estimated value = 0.66) between house price and income can therefore be taken as a good approximation for identifying sub-markets based on house prices.

first stage, an instrument variable was constructed (spatially lagged independent variable) and in the second stage, a quantile regression analysis was used with dependant variable of first equation as instruments.

4.2. Quantile regression model without spatial effects 4.3. Two stage quantile regression (2SQR) analysis of DID model Results for the quantile regression without the spatial effect are presented in Table 3. The second column shows the OLS estimation of the DID model for comparison purpose. Most variables are statistically significant with expected signs. Each quantile had reasonably good pseudo R2 values. One consistent finding is that most of the coefficients vary across quantiles and are different from the OLS estimation. Some variables reveal an increasing pattern in the values of coefficients across quantiles. For example, the price premium due to one additional bedroom is higher for high-value properties than for low-value property. Similar effects for bathrooms, gardens and waterfront location are also predicted. The variables D1 (impact of release of flood maps in 2009) and D2 (impact of actual flood in 2011) also show differences across different sub-markets which are discussed in our final model – the 2SQR estimation of DID. Following the OLS estimation and quantile regression analysis, spatial dependency was tested using Moran’s I statistics and which indicated the presence of spatial interaction. By employing the robust Lagrangian multiplier (LM) statistics, the presence of a spatial lag effect was further confirmed.6 2SQR analysis was then carried out in which the spatial lag variable was used as an instrument to correct for spatial interaction in the property market. In the

6 Moran’s I statistics = −0.432 (P-value = 0.01) and LM Lag statistics = 44.787 (Pvalue = 0.000).

A summary of the 2SQR results are presented in Table 4 which are comparable with quantile regression results discussed in the previous section. However, the coefficients are different as it is adjusted for the spatial lagged effect. The spatial coefficient is significant for most of ␶ values indicating the presence of a spatial lag effect. Of particular interest is the degree of variation exhibited by coefficient estimates across the distribution of house prices. As house prices increase (when the ␶ value is increasing), the positive impact of most structural variables increases. The flood variable is significant for most quantiles and has a negative effect on property prices varying between 4% and 8% across different sub-markets. Properties in flood-risk areas are discounted compared to properties outside flood risk areas. After release of flood risk information, property prices did not show a negative trend overall for the property market (BFAF INFOR) although the price increase is less in high-value properties. The first treatment – the property prices in flood risk areas after releasing of flood risk information (D1) – indicate a negative impact on property prices. However, after the 2011 flood, property prices show a decreasing trend in which the negative effect is higher for high-value properties. The estimated coefficients for D1 and D2 are different from OLS estimations as well as for those of different quantiles (see, Table 4). For instance, the coefficient for D1 is not significant for OLS (see Table 3) but shows a negative effect for high value flood affected properties. This suggests the release of flood risk

D. Rajapaksa et al. / Land Use Policy 62 (2017) 151–158

157

Table 4 Spatial quantile regression analysis (2SQR). Quantiles CONST BEDRM BATHRM GARAGE FLOORAREA HOUSETYPE FLOOR CARP FLOOR TILE FLOOR WOOD ROOF TILE ROOF ASBES WALL BRICK WALL WOOD POOL GARDEN STORIES WATERFRONT INCOME DIS PARK DIS HIGWAY DIS SHOP DIS RAIL DIS CBD DIS RIVER FLOOD BFAF INFOR BFAF FLOOD D1 D2 WY N

0.1 9.30998 −0.00637 0.09879**** 0.03436*** 0.0002625*** −0.01938 −0.00764 −0.11218 0.04052** −0.00557 −0.01720 −0.03762 0.00517 0.11944**** 0.10414**** 0.05692* 0.22117*** 0.000171**** 0.00000 −0.00002 −0.00004 0.00002 −0.00007**** −0.00001 −0.07303** 0.11809**** −0.00924 0.06159 −0.13686*** 0.28187 2345

0.25 10.64142** 0.02584*** 0.09254**** 0.040052**** 0.00025**** 0.01203* −0.00348 −0.00778 0.00799 −0.01861 −0.00838 −0.01180 0.03174 0.10846**** 0.08517**** 0.07992**** 0.32399**** 0.00017**** 0.00000 −0.00002 −0.00002 −0.00004** −0.00008**** 0.00003964* −0.04988** 0.07256**** −0.07875**** 0.01173 −0.10900*** 0.17154 2345

0.5 28.07864**** 0.05602**** 0.09519**** 0.02555**** 0.00036**** 0.02019**** −0.00870 0.01075 0.01478 −0.02612*** 0.00419 −0.04615**** −0.01926 0.10051**** 0.05783**** 0.08015**** 0.44826**** 0.00015**** −0.00005*** −0.00004*** −0.00001 −0.00006*** −0.00011**** 0.00008**** −0.07247**** 0.04137*** −0.0933**** −0.00732 −0.11095**** −1.11716*** 2345

0.75 42.76786**** 0.05410**** 0.12162**** 0.03166**** 0.00050**** 0.02036*** −0.00957 −0.01916 −0.00046 −0.03475*** 0.01055 −0.0509*** −0.03479** 0.13956**** 0.05260**** 0.10066**** 0.56754**** 0.00013**** −0.000062*** −0.00004** −0.00001 −0.00006*** −0.00013**** 0.00009**** −0.07658**** 0.02246* −0.11310**** −0.05594** −0.14505**** −2.21186**** 2345

0.9 45.12301**** 0.05298**** 0.13703**** 0.03357**** 0.00056**** 0.02405*** −0.02864 −0.03321 −0.02425 −0.04364*** −0.00275 −0.02624 0.00813 0.16700**** 0.05351**** 0.10542**** 0.53285**** 0.00014**** −0.00004 −0.00003* 0.00001 −0.00003 −0.00011**** 0.00003 −0.08144*** 0.03053 −0.12166**** −0.06217* −0.15188**** −2.4022**** 2345

Note: ****, ***, ** and * denotes coefficients are significant at 1,5,10 and 20% levels of significance, respectively.

information caused a more pronounced negative impact on highvalue properties than low-value properties in flood affected zones. According to D2, the negative impact of flood on flood affected properties is higher in middle range when compared with highvalue and low-value property. Gayer (2000) came to a similar conclusion in examining the impact of hazardous waste on property values. In a recent study Zhang (2016) also found that differences in property market behaviour across different property prices. Quantile regression analysis of DID models confirms the presence of property sub-markets. The Breusch-Pagan test also provides evidence of the presence of heteroskedasticity which justifies the estimation of quantile regression analysis.7 Using this means of submarket analysis, it can be shown that with the release of flood maps the property value for high-value sub-markets is decreased while the impact is insignificant for low-value properties. However, the second treatment (2011 flood) shows a significant negative impact which varies across sub-markets. As the results indicate, the negative impact is higher for both low-value and high-value properties compared with the middle group. In this way the results confirm the impact from natural disasters such as floods vary across different sub-markets. 5. Conclusion In spite of growing concern about the impact of floods on property values, the literature generally does not consider price variation across different sub-markets. Given this important gap in the flood impact literature, the main objective of this study is to examine the behaviour of different property sub-markets in terms of environmental amenities and dis-amenities.

7

Breusch-Pagan test: Chi2 = 182.92 (P-value = 0.000).

An important observation about HP analysis is that results vary according to the study area. In a review by Sirmans et al. (2005), only 21 of the 40 studies showed the expected sign for ‘number of bedrooms’ having a significant impact on property values. This lack of consistency of expected impact on property values was also evident for variables such as the distance to the CBD which displayed both positive and negative results. The inconsistency of such results confirms that HP analyses are specific to the selected research area and property sub-markets (Freeman, 1993). This equally indicates the validity of quantile regression estimation of HP price functions for sub-markets. Among many other advantages, traditional meanbased regression analysis tends to be outlier biased. That is, the mean price of properties does not generally represent a typical house given high priced properties result in an asymmetric distribution. Usefully, the quantile regression approach reduces the effect of outliers. In this study we assume the presence of property sub-markets based on property values. Buyers and sellers of high-value properties are considered as high-income groups, whereas, buyers and sellers of low-value properties are considered as low-income sub-markets. Members of the low-income category typically have limited access to bank loans and hence tend to consider only affordable low quality properties with fewer desired amenities. As those in the high income category have easier access to financial resources, they tend to buy high-value properties (Zietz et al., 2008). Therefore, the property value is considered to be a valid criterion for identifying property sub-market behaviour. The results of this study reconfirm the rich explanatory outcomes which can be derived from sub-market analysis based on property market price. By demonstrating the impact of treatment effects across different property groups it is shown that property values decrease in higher quantiles in response to the first treatment effect. The second treatment effect (2011 flood) reveals, on average, a decrease in

158

D. Rajapaksa et al. / Land Use Policy 62 (2017) 151–158

property prices of lower and higher quantiles compared to middle quantiles. Consequently, an important conclusion is that the negative effect of floods is higher for both low and high income groups (low and high-value properties). This study demonstrates an innovative approach in employing the existing methodology and differentiating it from the extant flood assessment studies. First, the results are discussed across different sub-markets and second, a DID estimation is extended to observe the treatment effect across different sub-markets. The results of this study have clear policy implications. They include identifying the scope for implementation of better calibrated insurance and flood mitigation programs as they relate to high income and low income property sub-markets. It is noted in this context that residents of a high-income and high-property value submarkets are in a position to afford extra insurance coverage and hence, despite the flood-risk, are more likely to continue to live in the area. However, the insurance premium may not be affordable for residents of the low-income sub-market. The issue of insurance is, therefore, one which bears further investigation including whether the differences in insurance policies between submarkets are justifiable and adequate (see, for example, Harrison et al., 2001). References Anselin, L., 1988. Spatial Econometrics: Methods and Models. Kluwer Academic Publishers, Dordrecht, The Netherlands. Bin, O., Landry, C.E., 2013. Changes in implicit flood risk premiums: empirical evidence from the housing market. J. Environ. Econ. Manage. 65 (3), 361–376. Bourassa, S.C., Hamelink, F., Hoesli, M., MacGregor, B.D., 1999. Defining housing submarkets. J. Hous. Econ. 8 (2), 160–183. Chernozhukov, v., Hansen, C., 2008. Instrumental variable quantile regression: a robust inference approach. J. Econometr. 142 (2), 379–398. Dale-Johnson, D., 1982. An alternative approach to housing market segmentation using hedonic price data. J. Urban Econ. 11 (3), 311–332. Dorsey, R.E., Hu, H., Mayer, W.J., Wang, H., 2010. Hedonic versus repeat-sales housing price indexes for measuring the recent boom-bust cycle. J. Hous. Econ. 19 (2), 75–93. Duarte, C.M., Tamez, C.G., 2009. Does noise have a stationary impact on residential value? J. Eur. Real Estate 2 (3), 259–279. Ebru, C., Eban, A., 2011. Determinants of house prices in Istanbul: a quantile regression approach. Qual. Quant. 45 (1), 305–317. Eves, C., Wilkinson, S., 2014. Assessing the immediate and short-term impact of flooding on residential property participant behaviour. Nat. Hazards 71 (3), 1519–1536. Farmer, M.C., Lipscomb, C.A., 2010. Using quantile regression in hedonic analysis to reveal submarket completion. J. Real Estate Res. 32 (4), 435–460. Filippova, O., Rehm, M., 2011. The impact of proximity to cell phone towers on residential property values. Int. J. Hous. Markets Anal. 4 (3), 244–267. Freeman, A.M., 1993. The Measurement of Environmental and Resource Values: Theory and Methods. Resources for the Future., Washington, DC. Fridgen, P.M., Shultz, S.D., 1999. The Influence of the Threat of Flooding on Housing Values in Fargo, North Dakota and Moorhead, Minnesota A Gricultural Economics Report. 417 Department of Agricultural Economics, North Dakota State University, Fargo. Gawande, K., Smith, H.J., Yuan, M., 2013. The long-run impact of nuclear waste shipments on the property market: evidence from a quasi-experiment. J. Environ. Econ. Manage. 65 (1), 56–73. Gayer, T., 2000. Neighbourhood demographics and the distribution of Hazardous Waste risks: an instrumental variables estimation. J. Regul. Econ. 17 (2), 131. Gopalakrishnan, S., Smith, M.D., Slott, J.M., Murray, A.B., 2011. The value of disappearing beaches: a hedonic pricing model with endogenous beach width. J. Environ. Econ. Manage. 61 (3), 297–310. Hao, L., Naiman, D.Q., 2007. Quantile Regression Sage University Papers Series on Quantitative Applications in Social Sciences, Series No. 149. Sage. Harrison, D.M., Smersh, G.T., Schwartz, A.L., 2001. Environmental determinants of housing prices: the impact of flood zone status. J. Real Estate Res. 21 (1/2), 3–20.

Heintzelman, M.D., 2010. Measuring the property –value effect of local land use and preservation referenda. Land Econ. 86 (1), 22–47. Hibiki, A., Managi, S., 2011. Does the housing market respond to information disclosure?: Effects of toxicity indices in Japan. J. Environ. Manage. 92 (1), 165–171. Hoshino, T., 2011. Estimation and analysis of preference heterogeneity in residential choice behaviour. Urban Stud. 48 (2), 363–382. Irwin, E.G., 2002. The effect of open space on residential property values. Land Econ. 78 (4), 465–480. Khandker, S.R., Koolwal, G.B., Samad, H.A., 2010. Handbook on Impact Evaluation: Quantitative Methods and Practices. World Bank Publications. Kim, T.H., Muller, C., 2004. Two-stage quantile regression when the first stage is based on quantile regression. Econometr. J. 7 (1), 218–231. Kim, C.W., Phipps, T.T., Anselin, L., 2003. Measuring the benefit of air quality improvement: a spatial hedonic approach. J. Environ. Econ. Manage. 45 (1), 24–39. Koenker, R., Bassett, G., 1978. Regression quantiles. Econometrica 46 (1), 33–50. Koenker, R., 2005. Quantile Regression. Cambridge University Press Cambridge. Kong, F., Yin, H., Nakagoshi, N., 2007. Using GIS landscape metrics in the hedonic price modelling of the amenity value of urban green space: a case study in Jinan city, China. Landscape Urban Plann. 79 (4), 240–252. Kostov, P., 2009. A spatial quantile regression hedonic model of agricultural land prices’. Spatial Econ. Anal. 4 (1), 53–72. Kuethe, T.H., Keeney, R., 2012. Environmental externalities and residential property values: externalized costs along the house price distribution. Land Econ. 88 (2), 241–250. Lamond, J., Proverbs, D., Hammond, F., 2010. The impact of flooding on the price of residential property: a transactional analysis of UK market. Housing Stud. 25 (3), 335–356. Landry, C.E., Hindsley, P., 2011. Valuing beach quality with hedonic property models. Land Econ. 87 (1), 92–108. Mak, S., Choy, L., Ho, W., 2010. Quantile regression estimation of Hong Kong real estate prices. Urban Stud. 47 (11), 2461–2472. McCluskey, J.J., Rausser, G.C., 2003. Stigmatized asset value: is it temporary or log-term? Rev. Econ. Stat. 85 (2), 276–285. McMillen, D., Thorsnes, P., 2006. Housing market and quantile repeat-sales price index. Real Estate Econ. 34 (4), 567–584. Merz, B., Kreibich, H., Thieken, A., Schwarze, R., 2004. Estimation uncertainty of direct monetary flood damage to buildings. Nat. Hazards Earth Syst. Sci. 4 (1), 153–160. Michaels, R.G., Smith, V.K., 1990. Market segmentation and valuing amenities with hedonic models: the case of hazardous waste sites. J. Urban Econ. 28 (2), 223–242. Miron, J.R., 1995. Place to place rent comparison among Canadian Cities. Geogr. Anal. 27 (2), 116–136. Nicholls, S., Crompton, J.L., 2005. The impact of greenways on property values: evidence from Austin, Texas. J. Leisure Res. 37 (3), 321–341. Petrolia, D.R., Landry, C.E., Coble, K.H., 2013. Risk preferences, risk perceptions and flood insurance. Land Econ. 89 (2), 227–245. Prasad, N., Richards, A., 2008. Improving median price indexes through stratification. J. Real Estate Res. 30 (1), 45–71. Rajapaksa, D., Wilson, C., Managi, S., Hoang, V., Lee, B., 2016. Flood risk information, actual floods and property values: a quasi-experimental analysis. Econ. Record 92, 52–67 (S1). Rambaldi, A.N., Fletcher, C.S., Collins, K., McAllister, R.R.J., 2013. Housing shadow prices in an inundation prone suburb. Urban Stud. 50 (9), 1889–1905. Samarasinghe, O., Sharp, B., 2010. Flood prone risk and amenity value: a spatial hedonic analysis. Aust. J. Agric. Res. Econ. 54 (4), 457–475. Sirmans, G.S., Macpgerson, D.A., Zietz, E.N., 2005. The composition of hedonic pricing models. J. Real Estate Literat. 13 (1), 3–43. Su, L., Yang, Z., 2007. Instrumental variable quantile estimation of spatial autoregressive models. Res. Coll. School Econ. (Available at:) http://ink.library. smu.edu.sg/soe research/1074. Thompson, E., Butters, R.B., Schmitz, B.T., 2012. The property value premium of a place of worship. Contemporary Econ. Policy 30 (2), 215–222. Wilhelmsson, M., 2004. A method to derive housing sub-markets and reduce spatial dependency. Property Manage. 22 (4), 276–288. Zabel, J.E., Kiel, K.A., 2000. Estimating the demand for air quality in four US cities. Land Econ. 76 (2), 174–194. Zhai, G., Fukuzono, T., 2003. Effect of flooding on Megalopolitan land price: a case study of the 2000 Tokai Flood in Japan. J. Nat. Disaster Sci. 25 (1), 23–36. Zhang, L., 2016. Flood hazards impact on neighborhood house prices: a spatial quantile regression analysis. Reg. Sci. Urban Econ. 60 (1), 12–19. Zietz, J., Zietz, E.N., Sirmans, G.S., 2008. Determinants of house price: a quantile regression approach. J. Real Estate Finance Econ. 37 (4), 317–333.